arXivSub Start free trial

ICML 2025 Papers with Code β€” Page 5

International Conference on Machine Learning Β· 722 papers

Long-Form Speech Generation with Spoken Language Models

Se Jin Park (Google DeepMind), RJ Skerry-Ryan (Google DeepMind)

CodeGenerationBenchmarkAudio

🎯 What it does: This paper presents SpeechSSMβ€”a text-independent speech language model based on state space models, capable of generating several minutes of continuous speech in one go.

Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models

Xin Zou (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)

CodeGenerationRetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodalityRetrieval-Augmented Generation

🎯 What it does: This paper proposes MemVR, a decoding strategy based on visual memory re-inspection that dynamically triggers visual re-inspection during the generation process of multimodal large language models, enhancing factual consistency and reducing hallucinations.

Low-Rank Tensor Transitions (LoRT) for Transferable Tensor Regression

Andong Wang (RIKEN AIP), Qibin Zhao (RIKEN AIP)

CodeRestorationDomain AdaptationComputational EfficiencyVideo

🎯 What it does: This paper proposes the Low-Rank Tensor Transitions (LoRT) framework, aimed at addressing the tensor regression problem in scenarios of insufficient samples, model/covariate shift, and decentralized data environments; a distributed version, D-LoRT, is also provided.

LV-XAttn: Distributed Cross-Attention for Long Visual Inputs in Multimodal Large Language Models

Tzu-Tao Chang (University of Wisconsin-Madison), Shivaram Venkataraman (University of Wisconsin-Madison)

CodeComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoMultimodality

🎯 What it does: A distributed and precise cross-attention mechanism (LV-XAttn) and activation recomputation technique are proposed for efficiently processing long visual inputs in multimodal large language models.

MA-LoT: Model-Collaboration Lean-based Long Chain-of-Thought Reasoning enhances Formal Theorem Proving

Ruida WANG, Tong Zhang (University of Illinois Urbana Champaign)

CodeLarge Language ModelTextChain-of-Thought

🎯 What it does: The MA-LoT framework is proposed, which splits the Lean4 formal proof task into two stages: global proof planning and fine-grained error correction, and trains a large language model (LLM) with long chain reasoning (Long CoT) capability through LoT-Transfer Learning.

Machine Learning meets Algebraic Combinatorics: A Suite of Datasets Capturing Research-level Conjecturing Ability in Pure Mathematics

Herman Chau (University of Washington), Henry Kvinge (Pacific Northwest National Laboratory)

CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphTabular

🎯 What it does: This paper proposes the 'ACD Data Warehouse' aimed at combinatorial algebra, which constructs nine categories of datasets containing a large number of instances and open research questions, with the goal of inspiring machine learning models during the conjecture phase;

Making Hard Problems Easier with Custom Data Distributions and Loss Regularization: A Case Study in Modular Arithmetic

Eshika Saxena (Meta AI), Kristin E. Lauter

CodeTransformer

🎯 What it does: The research improves the training data distribution and loss function, allowing machine learning models to achieve higher accuracy in modular arithmetic tasks (especially large-scale modular summation) and in attacks on Learning With Errors (LWE).

MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models

Mahir Labib Dihan (Bangladesh University of Engineering and Technology), Md Rizwan Parvez (Qatar Computing Research Institute)

CodeTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark

🎯 What it does: The MapEval benchmark is proposed, which evaluates the geospatial reasoning capabilities of large language models and vision-language models in three scenarios: text, API interaction, and visual maps, using 700 multiple-choice questions.

Mask-Enhanced Autoregressive Prediction: Pay Less Attention to Learn More

Xialie Zhuang (University of Chinese Academy of Sciences), Shiwei Liu (University of Oxford)

CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: Without changing the existing decoder-only structure, 10%-15% of the input tokens are randomly masked, and then standard next token prediction is directly performed, constructing the Mask-Enhanced Autoregressive Prediction (MEAP) training paradigm.

Masked Autoencoders Are Effective Tokenizers for Diffusion Models

Hao Chen (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)

CodeGenerationData SynthesisTransformerDiffusion modelAuto EncoderImage

🎯 What it does: Proposes MAETok, using a masked autoencoder as a tokenizer to construct a structured latent space to enhance the generation quality of diffusion models.

MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation

Jiawen Wang (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)

CodeSegmentationDomain AdaptationImageBiomedical Data

🎯 What it does: The MaskTwins framework is proposed, which enforces consistency learning on target domain images through bidirectional complementary masks, achieving unsupervised domain adaptive semantic segmentation.

Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding

Mingyu Jin (Rutgers University), Yongfeng Zhang (Rutgers University)

CodeTransformerLarge Language ModelText

🎯 What it does: This paper systematically analyzes the 'massive values' generated by the attention module's Q and K in large language models, revealing their key role in understanding contextual knowledge.

MATS: An Audio Language Model under Text-only Supervision

Wen Wang (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)

CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityAudio

🎯 What it does: This paper presents MATS, a multimodal large language model capable of performing various audio tasks (classification, subtitles, question answering) using only text supervision.

Maximum Entropy Reinforcement Learning with Diffusion Policy

Xiaoyi Dong (Institute of Automation, Chinese Academy of Sciences), Xi Sheryl Zhang (University of Chinese Academy of Sciences)

CodeReinforcement LearningDiffusion modelTabularOrdinary Differential Equation

🎯 What it does: This paper proposes the use of diffusion models as policy representations within the maximum entropy reinforcement learning framework and provides methods for training and probability estimation.

MCU: An Evaluation Framework for Open-Ended Game Agents

Xinyue Zheng (Beijing Institute for General Artificial Intelligence), Yitao Liang (Peking University)

CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality

🎯 What it does: An open game agent evaluation framework named MCU was built in Minecraft, which includes 3,452 combinable atomic tasks, a task combination mechanism, and an automatic evaluation system based on visual-language models;

MedXpertQA: Benchmarking Expert-Level Medical Reasoning and Understanding

Yuxin Zuo (Tsinghua University), Bowen Zhou (Tsinghua University)

CodeTransformerLarge Language ModelTextMultimodalityBenchmark

🎯 What it does: Proposes MedXpertQA, an expert-level medical reasoning and understanding benchmark covering 17 departments, 11 systems, and 4,460 questions, including text and multimodal subsets;

MERIT: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training

Yang Luo (National University of Singapore), Yang You (National University of Singapore)

CodeOptimizationTransformerLarge Language ModelText

🎯 What it does: The MERIT optimizer is proposed, which enhances the stability and effectiveness of training language models at large batch sizes by utilizing maximum norm and element-wise trust ratios.

MetaAgent: Automatically Constructing Multi-Agent Systems Based on Finite State Machines

Yaolun Zhang (University of Wisconsin Madison), Chaowei Xiao (University of Wisconsin Madison)

CodeLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Automatically construct a multi-agent system framework called MetaAgent, which uses finite state machines to automatically generate and optimize multi-agent processes based on task descriptions.

MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters

Arsalan Sharifnassab (Openmind Research Institute), Richard S. Sutton (University of Alberta)

CodeOptimizationHyperparameter SearchConvolutional Neural NetworkTransformerImageText

🎯 What it does: A general framework called MetaOptimize is proposed, which can optimize hyperparameters such as learning rate in real-time during the training process to minimize the regret of discounted future loss.

Mind the Gap: A Practical Attack on GGUF Quantization

Kazuki Egashira (ETH Zurich), Martin Vechev (ETH Zurich)

CodeOptimizationAdversarial AttackText

🎯 What it does: An attack framework for the GGUF quantization method (k-quants) is proposed, which can inject malicious behavior after model quantization while the full-precision version remains normal.

Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers

Thiziri Nait Saada (Mathematical Institute, University of Oxford), Jared Tanner (Mathematical Institute, University of Oxford)

CodeTransformer

🎯 What it does: Analyzed the spectral properties of the softmax self-attention layer in the Transformer during random initialization, revealing rank collapse in the width direction (tokens converge to a single representation within a single layer) and issues leading to gradient explosion.

MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data

Yuqin Dai (Nanjing University of Science and Technology), Jiamin Wu (Chinese University of Hong Kong)

CodeRetrievalExplainability and InterpretabilityContrastive LearningBiomedical DataMagnetic Resonance Imaging

🎯 What it does: We propose MindAligner, an explicit brain function alignment framework that maps the fMRI signals of new subjects to the known subject space through a brain transfer matrix, enabling cross-subject visual decoding.

MiraGe: Editable 2D Images using Gaussian Splatting

Joanna Waczynska, PrzemysΕ‚aw Spurek (Jagiellonian University)

CodeRestorationGenerationGaussian SplattingImage

🎯 What it does: Reconstructing and editing two-dimensional images in three-dimensional space using planar three-dimensional Gaussian distribution;

Mitigating Over-Squashing in Graph Neural Networks by Spectrum-Preserving Sparsification

Langzhang Liang (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)

CodeOptimizationGraph Neural NetworkNeural Radiance FieldGraph

🎯 What it does: This paper proposes a graph reconnection method called GOKU based on spectral-preserving sparsification to alleviate the over-compression problem in graph neural networks.

Mitigating Plasticity Loss in Continual Reinforcement Learning by Reducing Churn

Hongyao Tang (Mila Quebec AI Institute), Glen Berseth (Mila Quebec AI Institute)

CodeReinforcement LearningSequential

🎯 What it does: This paper explains and addresses the issue of 'plasticity loss' in long-term learning, which refers to the network's loss of adaptability to new tasks, by studying 'churn' (the fluctuation of network outputs on untrained samples) in deep continual reinforcement learning.

MixBridge: Heterogeneous Image-to-Image Backdoor Attack through Mixture of SchrΓΆdinger Bridges

Shixi Qin (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

CodeGenerationData SynthesisAdversarial AttackMixture of ExpertsDiffusion modelImageStochastic Differential Equation

🎯 What it does: This paper proposes MixBridge, an image-to-image diffusion model based on the Schrâdinger bridge, which simultaneously achieves clean generation and various hidden attacks within the same model.

Mixed-curvature decision trees and random forests

Philippe Chlenski (Columbia University), Itsik Pe'er (Columbia University)

CodeClassificationOptimizationTabularTime Series

🎯 What it does: A general framework for decision tree and random forest learning on mixed curvature product manifolds is proposed, utilizing angular projection in two-dimensional subspaces to achieve decision splits;

Mixture of Experts Made Intrinsically Interpretable

Xingyi Yang (University of Oxford), Philip Torr (University of Oxford)

CodeExplainability and InterpretabilityTransformerMixture of ExpertsText

🎯 What it does: An interpretable mixture of experts language model MoE-X is proposed, which eliminates multi-semantic neurons by using ReLU and sparse routing among experts.

Mixture of Lookup Experts

Shibo Jie (Peking University), Yunhe Wang (Huawei)

CodeMixture of ExpertsText

🎯 What it does: This paper proposes the Mixture of Lookup Experts (MoLE), which reparameterizes expert parameters as a lookup table (LUT), allowing for inference without loading experts, thereby significantly reducing memory usage and latency.

MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization

Kangyu Zhu (University of North Carolina at Chapel Hill), Huaxiu Yao (University of North Carolina at Chapel Hill)

CodeOptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBiomedical Data

🎯 What it does: Proposes MMedPO, a multimodal preference optimization method that uses clinical relevance as a weight to enhance the factuality of medical visual language models by generating preference samples with hallucinations and local lesion noise.

Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence

Shangbin Feng (University of Washington), Tomas Pfister (Google)

CodeOptimizationTransformerLarge Language ModelText

🎯 What it does: Proposes MODEL SWARMS, a collaborative search algorithm based on particle swarm optimization, designed for adaptive large language models in low data scenarios;

Model-Based Exploration in Monitored Markov Decision Processes

Alireza Kazemipour (University of Alberta), Michael Bowling (University of Alberta)

CodeOptimizationReinforcement LearningTabularBenchmark

🎯 What it does: This paper proposes a new model-based algorithm Monitored MBIE-EB to address the partially observable reward Monitored Markov Decision Process (Mon-MDP) problem, and provides the first theoretical guarantee for finite sample complexity.

Modified K-means Algorithm with Local Optimality Guarantees

Mingyi Li (University of Tokyo), Akiko Takeda (University of Tokyo)

CodeOptimizationTabular

🎯 What it does: The theoretical analysis of the local optimality of the K-means algorithm is conducted, providing counterexamples to prove that traditional assumptions do not hold, and a lightweight variant called LO-K-means is proposed, which guarantees convergence to continuous or discrete local optimal solutions while maintaining the same time complexity.

MOGIC: Metadata-infused Oracle Guidance for Improved Extreme Classification

Suchith Chidananda Prabhu (Indian Institute of Technology Delhi), Manik Varma (Microsoft)

CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The MOGIC framework is proposed, employing a two-stage training approach that utilizes an oracle containing metadata from both the query side and the label side to guide a low-latency extreme classification model, achieving higher prediction accuracy.

MoH: Multi-Head Attention as Mixture-of-Head Attention

Peng Jin (Peking University), Shuicheng YAN

CodeClassificationGenerationComputational EfficiencyTransformerMixture of ExpertsImage

🎯 What it does: Proposes Mixture-of-Head Attention (MoH), which dynamically routes each token to activate only a portion of the heads in the multi-head attention of the Transformer and replaces standard summation with weighted summation, significantly reducing computational load;

MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking

Sebastian Farquhar (Google DeepMind), Rohin Shah (Google DeepMind)

CodeOptimizationReinforcement LearningTabular

🎯 What it does: This paper proposes and validates a method that combines myopic optimization with foresight approval (MONA) to prevent reinforcement learning agents from performing multi-step reward hacking without being detected by supervisors.

Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models

Chao Li (Intel Labs China), Anbang Yao (Intel Labs China)

CodeGenerationData SynthesisComputational EfficiencyDiffusion modelImage

🎯 What it does: In this work, the authors propose a dual-sampling framework named Morse, which achieves lossless acceleration of the generative process of diffusion models by combining jump sampling with residual feedback.

Multi-band Frequency Reconstruction for Neural Psychoacoustic Coding

Dianwen Ng (Nanyang Technological University), EngSiong Chng

CodeCompressionConvolutional Neural NetworkAuto EncoderAudio

🎯 What it does: This paper proposes and implements MUFFIN, a fully convolutional neural psychoacoustic encoder based on multi-band residual vector quantization, achieving high-fidelity compression across different audio types.

Multi-Objective Causal Bayesian Optimization

Shriya Bhatija (Technical University of Munich), Thomas BohnΓ© (University of Cambridge)

CodeOptimizationGraphTabularBiomedical DataFinance Related

🎯 What it does: Proposes Multi-Objective Causal Bayesian Optimization (MO-CBO) to find Pareto optimal intervention combinations under known causal graphs.

Multivariate Conformal Selection

Tian Bai (McGill University), Archer Y. Yang (McGill University)

CodeDrug DiscoveryTabularBiomedical DataAlzheimer's Disease

🎯 What it does: A multivariate Conformal Selection (mCS) method is proposed, extending the original CS to the case of multidimensional responses, utilizing regional monotonicity to ensure finite sample FDR control, and achieving efficient candidate selection through two types of inconsistency scores.

Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer

Yabo Liu (Harbin Institute of Technology), Jinghua Wang

CodeObject DetectionSegmentationDomain AdaptationKnowledge DistillationSupervised Fine-TuningImageMultimodality

🎯 What it does: In the source unsupervised domain transfer scenario, a dual-network collaborative adaptive framework is designed for the Segment Anything Model (SAM) to achieve zero-label transfer to the target domain.

MxMoE: Mixed-precision Quantization for MoE with Accuracy and Performance Co-Design

Haojie Duanmu (Shanghai Jiao Tong University), Dahua Lin (The Chinese University of Hong Kong)

CodeMixture of ExpertsText

🎯 What it does: This paper proposes and implements MxMoE, a mixed-precision quantization framework for Mixture-of-Experts models, aimed at balancing model accuracy and inference performance.

Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning

Fangwen Wu (Zhejiang Lab), Meng Wang (Hefei University of Technology)

CodeClassificationKnowledge DistillationTransformerSupervised Fine-TuningImage

🎯 What it does: In category incremental learning unrelated to research tasks, semantic drift caused by low-rank adaptation is addressed by proposing two calibration methods: mean shift compensation and covariance calibration, combined with feature-level self-distillation to improve model stability.

NBDI: A Simple and Effective Termination Condition for Skill Extraction from Task-Agnostic Demonstrations

Myunsoo Kim (Korea University), Byung-Jun Lee (Gauss Labs Inc.)

CodeRobotic IntelligenceReinforcement LearningSequential

🎯 What it does: This paper proposes a decision point identification method based on state-action novelty, NBDI, which can learn terminable skills from task-agnostic demonstrations.

Near-Optimal Decision Trees in a SPLIT Second

Varun Babbar (Duke University), Margo Seltzer (University of British Columbia)

CodeClassificationOptimizationComputational EfficiencyTabular

🎯 What it does: The SPLIT series algorithms (SPLIT, LicketySPLIT, and RESPLIT) are proposed, which achieve approximately optimal decision trees by using dynamic programming + branch-and-bound search in the shallow layers of the tree and greedy splitting in the deeper layers, allowing for the rapid generation of the Rashomon set.

Near-optimal Sketchy Natural Gradients for Physics-Informed Neural Networks

Maricela Best Mckay, Brian Wetton (University of British Columbia)

CodeOptimizationComputational EfficiencyTabularPhysics Related

🎯 What it does: A randomized natural gradient algorithm (SNGD) is proposed, which significantly reduces the computational cost and memory usage during training by compressing the Gram matrix of PINN.

NegMerge: Sign-Consensual Weight Merging for Machine Unlearning

Hyo Seo Kim (Sogang University), Junsuk Choe (Sogang University)

CodeClassificationData-Centric LearningTransformerVision Language ModelImage

🎯 What it does: A novel machine forgetting method called NegMerge is proposed, which effectively deletes specified knowledge in the model by merging task vectors from multiple models.

Nemotron-CORTEXA: Enhancing LLM Agents for Software Engineering Tasks via Improved Localization and Solution Diversity

Atefeh Sohrabizadeh (NVIDIA), Bryan Catanzaro (NVIDIA)

CodeAI Code AssistantTransformerLarge Language ModelContrastive LearningTextBenchmark

🎯 What it does: The NEMOTRON-CORTEXA system is proposed, utilizing a customized code embedding model, location agents based on AST and LSP, and diversified patch generation to enhance the efficiency of locating and fixing LLM software engineering tasks.

Neural Discovery in Mathematics: Do Machines Dream of Colored Planes?

Konrad Mundinger (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)

CodeOptimization

🎯 What it does: This paper proposes a method to transform the Hadwiger-Nelson problem into a continuous optimization task, utilizing neural networks for gradient search on a differentiable loss function, ultimately discovering a new six-color plane coloring scheme.

Neural Event-Triggered Control with Optimal Scheduling

Luan Yang (Fudan University), Wei Lin (Fudan University)

CodeOptimizationTime SeriesOrdinary Differential Equation

🎯 What it does: A neural network event-triggered control framework, Neural ETC, is proposed to achieve stable control of the system under limited communication resources.

Neural Genetic Search in Discrete Spaces

Hyeonah Kim (Quebec Artificial Intelligence Institute), Changhyun Kwon (KAIST)

CodeGenerationOptimizationLarge Language ModelTextGraph

🎯 What it does: This paper proposes a search method that integrates a genetic algorithm mechanism during the testing phase of deep generative modelsβ€”Neural Genetic Search (NGS).

Neural Solver Selection for Combinatorial Optimization

Chengrui Gao (Nanjing University), Chao Qian (Nanjing University)

CodeOptimizationGraph Neural NetworkGraph

🎯 What it does: A general framework is proposed to select the most suitable neural solver at the instance level to solve combinatorial optimization problems.

NeuronTune: Towards Self-Guided Spurious Bias Mitigation

Guangtao Zheng (University of Virginia), Aidong Zhang (University of Virginia)

CodeClassificationOptimizationImageTextMultimodality

🎯 What it does: A post-hoc unsupervised method called NeuronTune is proposed, which alleviates the model's dependence on spurious correlations by identifying and suppressing the neurons that produce bias within the embedded space of a trained model.

NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders

Jun-En Ding (Stevens Institute of Technology), Feng Liu (Stevens Institute of Technology)

CodeGraph Neural NetworkContrastive LearningGraphBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation

🎯 What it does: This paper proposes the NEUROTREE framework, which combines k-hop AGE-GCN, neural ODE, and Contrastive Masked Functional Connectivity (CMFC) to decode fMRI functional networks in a tree structure, enabling hierarchical analysis and interpretation of brain pathways in mental disorders.

NExtLong: Toward Effective Long-Context Training without Long Documents

Chaochen Gao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)

CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: The NExtLong framework is proposed, which extends long-context data through negative documents and trains LLMs;

No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets

Corinna Coupette (Aalto University), Bastian Rieck (University of Fribourg)

CodeData-Centric LearningGraph Neural NetworkGraph

🎯 What it does: Proposes the RINGS framework, which evaluates the quality of graph learning datasets through pattern perturbation (graph structure or node features); introduces two metrics: performance separability and pattern complementarity.

No Soundness in the Real World: On the Challenges of the Verification of Deployed Neural Networks

Attila SzΓ‘sz (University of Szeged), MΓ‘rk Jelasity (University of Szeged)

CodeAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper proves that theoretically claimable verifiable neural networks are not necessarily secure in real deployment environments, and demonstrates through the construction of backdoored networks that can be triggered in specific deployed environments that existing state-of-the-art verifiers cannot guarantee actual deployment security.

NoLiMa: Long-Context Evaluation Beyond Literal Matching

Ali Modarressi (Ludwig Maximilian University of Munich), Hinrich Schuetze

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought

🎯 What it does: Proposes the NOLIMA benchmark for evaluating the associative reasoning ability of LLMs beyond literal matching in long contexts.

Nonparametric Identification of Latent Concepts

Yujia Zheng (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

CodeFlow-based ModelImage

🎯 What it does: A non-parametric theoretical framework is proposed, demonstrating that when observations from different categories are sufficiently diverse, hidden concepts and their connection structures can be theoretically identified from the observational data.

Normalizing Flows are Capable Generative Models

Shuangfei Zhai (Apple), Joshua M. Susskind (Apple)

CodeGenerationData SynthesisTransformerScore-based ModelFlow-based ModelImage

🎯 What it does: This paper proposes a Transformer-based Block Autoregressive Flow architecture called TARFLOW, which constructs a powerful Normalizing Flow generative model through Gaussian noise augmentation, score-based denoising, and guidance techniques.

Novelty Detection in Reinforcement Learning with World Models

Geigh Zollicoffer (Georgia Institute of Technology), Mark Riedl (Georgia Institute of Technology)

CodeAnomaly DetectionReinforcement LearningWorld ModelSequential

🎯 What it does: This paper proposes a KL divergence upper bound based on a world model to automatically detect novelty in reinforcement learning environments without the need for manual thresholds.

NTK-DFL: Enhancing Decentralized Federated Learning in Heterogeneous Settings via Neural Tangent Kernel

Gabriel Thompson (North Carolina State University), Huaiyu Dai (North Carolina State University)

CodeFederated LearningImage

🎯 What it does: This paper proposes a decentralized federated learning framework called NTK-DFL, which drives model weight evolution through the Neural Tangent Kernel (NTK) and achieves global model learning by averaging parameters among neighbors.

Objective drives the consistency of representational similarity across datasets

Laure Ciernik (Technische UniversitΓ€t Berlin), Lukas Muttenthaler (Technische UniversitΓ€t Berlin)

CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: A comparison of the internal representations of 64 visual models across 23 different datasets is conducted, proposing a framework based on similarity vectors to measure the consistency of model representation similarity across different datasets.

Occult: Optimizing Collaborative Communications across Experts for Accelerated Parallel MoE Training and Inference

Shuqing Luo (University of North Carolina), Tianlong Chen (University of North Carolina)

CodeOptimizationComputational EfficiencyLarge Language ModelMixture of ExpertsText

🎯 What it does: An algorithm named Occult is proposed - a system-coordinated design scheme to optimize the all-to-all communication of experts in the Mixture-of-Experts (MoE) model, thereby accelerating the training and inference of MoE LLMs.

OmniBal: Towards Fast Instruction-Tuning for Vision-Language Models via Omniverse Computation Balance

Yongqiang Yao (Shanghai Jiao Tong University), Ningyi Xu (Shanghai Jiao Tong University)

CodeOptimizationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelMultimodality

🎯 What it does: Designed and implemented the OmniBal framework, which systematically balances the computational load of data, models, and memory, significantly accelerating the instruction fine-tuning training of VLM.

On Explaining Equivariant Graph Networks via Improved Relevance Propagation

Hongyi Ling (Texas A&M University), Shuiwang Ji (Texas A&M University)

CodeExplainability and InterpretabilityGraph Neural NetworkGraph

🎯 What it does: The EquiGX method is proposed, which implements layer-wise relevance propagation for spherical equivariant GNNs using deep Taylor decomposition, thereby providing an interpretable decomposition of the prediction results.

On the Diversity of Adversarial Ensemble Learning

Jun-Qi Guo (Nanjing University), Zhi-Hua Zhou (Nanjing University)

CodeAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage

🎯 What it does: This paper studies the concept of diversity in adversarial ensemble learning, proves that calculating exact diversity is NP-Hard, provides a first-order approximation-based diversity decomposition, and designs the AdvEOAP ensemble method to enhance gradient and cross diversity through orthogonal adversarial prediction.

On the Generalization Ability of Next-Token-Prediction Pretraining

Zhihao Li (Huazhong Agricultural University), Feng Zheng (Southern University of Science and Technology)

CodeTransformerLarge Language ModelText

🎯 What it does: This paper establishes a detailed generalization analysis of next token prediction (NTP) pre-training based on Rademacher complexity, exploring how NTP pre-training affects the model's generalization ability.

On the Out-of-Distribution Generalization of Self-Supervised Learning

Wenwen Qiang (Institute of Software Chinese Academy of Sciences), Changwen Zheng (Institute of Software Chinese Academy of Sciences)

CodeDomain AdaptationRepresentation LearningContrastive LearningImage

🎯 What it does: This paper conducts an in-depth analysis of the out-of-distribution (OOD) generalization problem in self-supervised learning (SSL) from the perspectives of batch construction and causality. It introduces the concept of post-intervention distribution (PID) and designs a batch sampling strategy based on latent variable models and propensity scores to eliminate spurious correlations among factors during the self-supervised training process, thereby enhancing SSL performance on OOD tasks.

One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning via Global Model Trajectory

Yuhang Li (Shanghai University), Xiaoqiang Li (Singapore Management University)

CodeOptimizationFederated LearningConvolutional Neural NetworkTransformerImageText

🎯 What it does: This paper studies the introduction of global model trajectories in federated learning, directly measuring and minimizing the sharpness of global loss, and proposes the FedGMT algorithm to alleviate the client drift problem caused by data heterogeneity, while reducing the computational cost of SAM through a single backward propagation.

One Example Shown, Many Concepts Known! Counterexample-Driven Conceptual Reasoning in Mathematical LLMs

Yinghui Li (Tsinghua University), Philip S. Yu (University of Illinois Chicago)

CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark

🎯 What it does: A university-level counterexample-driven conceptual reasoning benchmark, COUNTERMATH, has been constructed to evaluate the mathematical reasoning capabilities of LLMs.

One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework

Feiran Li (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

CodeGenerationData SynthesisDiffusion modelImageText

🎯 What it does: A text-image collaborative concept elimination framework (Co-Erasing) is proposed, which jointly guides the elimination of undesirable concepts in diffusion models through text prompts and corresponding image prompts.

Online Episodic Convex Reinforcement Learning

Bianca Marin Moreno (Univ. Grenoble Alpes), Nadia Oudjane (EDF)

CodeOptimizationReinforcement Learning

🎯 What it does: This paper proposes an online convex reinforcement learning framework (CURL) and presents an algorithm that achieves near-optimal asymptotic unbounded rewards under unknown transition dynamics and variable opponent losses.

Online Laplacian-Based Representation Learning in Reinforcement Learning

Maheed H. Ahmed (Purdue University), Mahsa Ghasemi (Purdue University)

CodeRepresentation LearningReinforcement LearningGraph

🎯 What it does: An online learning framework is proposed for synchronously updating policies and Laplacian representations in reinforcement learning, utilizing Asymmetric Graph Drawing Objectives (AGDO) for representation learning;

Online Sparsification of Bipartite-Like Clusters in Graphs

Joyentanuj Das (University of Edinburgh), He Sun (University of Edinburgh)

CodeOptimizationComputational EfficiencyGraph Neural NetworkGraph

🎯 What it does: This paper proposes an online sparsification algorithm that can construct sparse subgraphs in nearly linear time while preserving the structure and metrics of k bipartite-like clusters in undirected and directed graphs.

Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction

Yanbin Wei (Southern University of Science and Technology), Yu Zhang (Hong Kong University of Science and Technology)

CodeGraph Neural NetworkGraph

🎯 What it does: Proposed GVN (Graph Vision Network) and its efficient variant E-GVN for the link prediction task, utilizing graph visualization to generate images and extracting structural features through a visual encoder, which is then combined with MPNN.

OpenworldAUC: Towards Unified Evaluation and Optimization for Open-world Prompt Tuning

Cong Hua (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

CodeClassificationOptimizationTransformerPrompt EngineeringVision Language ModelImage

🎯 What it does: Proposed the OpenworldAUC metric and the Gated Mixture-of-Prompts (GMoP) framework, achieving a unified evaluation and optimization of Open-world Prompt Tuning.

Optimal Information Retention for Time-Series Explanations

Jinghang Yue (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)

CodeOptimizationExplainability and InterpretabilityContrastive LearningTime SeriesElectrocardiogram

🎯 What it does: This paper proposes an optimal information retention principle based on information theory and implements the ORTE framework for generating precise local explanations of temporal models.

Optimization Proxies using Limited Labeled Data and Training Time -- A Semi-Supervised Bayesian Neural Network Approach

Parikshit Pareek (Indian Institute of Technology Roorkee), Deepjyoti Deka (MIT Energy Initiative)

CodeOptimizationTabular

🎯 What it does: The study proposes a semi-supervised Bayesian neural network (Sandwich BNN) as an optimization surrogate, which quickly approaches the solution of constrained optimization problems using a small amount of labeled data and limited training time.

Optimizing Adaptive Attacks against Watermarks for Language Models

Abdulrahman Diaa (University of Waterloo), Nils Lukas (Mohammed Bin Zayed University of Artificial Intelligence)

CodeOptimizationAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: This paper addresses the robustness of text watermarking in large language models (LLMs) by proposing the construction of a preference dataset using observable watermark algorithms and proxy models. It employs reinforcement learning (such as DPO) to adaptively tune the rewriting model with open weights, enabling it to efficiently evade watermark detection while maintaining text quality.

Optimizing Large Language Model Training Using FP4 Quantization

Ruizhe Wang (University of Science and Technology of China), Peng CHENG

CodeOptimizationTransformerLarge Language ModelText

🎯 What it does: The first FP4 low-precision training framework for large language models (LLM) is proposed, demonstrating the feasibility of training a 13B parameter model from scratch with 100B tokens.

Optimizing Temperature for Language Models with Multi-Sample Inference

Weihua Du (Carnegie Mellon University), Sean Welleck (Carnegie Mellon University)

CodeOptimizationTransformerLarge Language ModelText

🎯 What it does: This paper studies the temperature selection problem under multi-sample aggregation strategies and proposes an automatic temperature optimization method based on entropy inflection points.

OR-Bench: An Over-Refusal Benchmark for Large Language Models

Justin Cui (University of California Los Angeles), Cho-Jui Hsieh (University of California Los Angeles)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Proposed and implemented an automated generation of a large-scale 'over-rejection' benchmark dataset OR-Bench, including 80K prompts, 1K difficult problems, and 600 toxic prompts.

Otter: Generating Tests from Issues to Validate SWE Patches

Toufique Ahmed (IBM Research), Martin Hirzel (IBM Research)

CodeGenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: The Otter system is proposed, which utilizes LLM combined with rule-driven analysis and self-reflective action planning to automatically generate fail-to-pass tests based on issue descriptions and old code before receiving a fix patch, supporting TDD and automated repair agents; and has released the TDD-Bench-Verified benchmark.

OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition

Zheng Lian (Institute of Automation, Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)

CodeRecognitionTransformerLarge Language ModelVideoTextMultimodalityAudio

🎯 What it does: Proposed and implemented the Open Vocabulary Multimodal Emotion Recognition (OV-MER) paradigm, constructed the OV-MERD dataset, designed evaluation metrics based on emotion clustering, and conducted benchmark experiments on multimodal large language models.

Pairwise Maximum Likelihood For Multi-Class Logistic Regression Model With Multiple Rare Classes

Xuetong Li (Xi'an Jiaotong University), Hansheng Wang (Peking University)

CodeClassificationOptimizationComputational EfficiencyImage

🎯 What it does: This paper studies the situation in multi-class logistic regression where there is one main class and several rare classes. It proposes estimation methods based on Pairwise Maximum Likelihood Estimation (PMLE) and Subsample Pairwise Maximum Likelihood Estimation (SPMLE), aiming to address the computational difficulties of traditional Global Maximum Likelihood Estimation (GMLE) in high-dimensional, large-class imbalanced data.

PAK-UCB Contextual Bandit: An Online Learning Approach to Prompt-Aware Selection of Generative Models and LLMs

Xiaoyan Hu (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)

CodeGenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageVideoText

🎯 What it does: An online learning framework for dynamically selecting generative models based on prompts, PAK-UCB and RFF-UCB, is proposed, which uses contextual bandits to adaptively select the best text/image/video generative model or LLM.

PANDAS: Improving Many-shot Jailbreaking via Positive Affirmation, Negative Demonstration, and Adaptive Sampling

Avery Ma (University of Toronto), Amir-massoud Farahmand (Polytechnique Montreal)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: The PANDAS method is proposed, which significantly improves the success rate of jailbreaks by incorporating Positive Affirmation (PA), Negative Demonstration (ND), and Adaptive Sampling (AS) targeting prompt themes in multi-shot jailbreaks.

Pareto-frontier Entropy Search with Variational Lower Bound Maximization

Masanori Ishikura (Nagoya Institute of Technology), Masayuki Karasuyama (Nagoya Institute of Technology)

CodeOptimizationHyperparameter SearchTabular

🎯 What it does: A Pareto Front Entropy Search Method based on Variational Lower Bound Maximization (PFEV) is proposed for multi-objective Bayesian optimization.

Pareto-Optimal Fronts for Benchmarking Symbolic Regression Algorithms

Kei Sen Fong (National University of Singapore), Mehul Motani (National University of Singapore)

CodeOptimizationTabularBenchmark

🎯 What it does: By using Gene Expression Programming (GEP) for exhaustive search on 34 black-box datasets from SRBench, an Absolute Pareto Front (APO front) was constructed, and a systematic comparison of the performance of 8 numerical optimization methods in obtaining the APO front was conducted; at the same time, general standards for benchmark evaluation and visualization were proposed.

Peri-LN: Revisiting Normalization Layer in the Transformer Architecture

Jeonghoon Kim (NAVERCloud), Kang Min Yoo (NAVER)

CodeTransformerLarge Language ModelText

🎯 What it does: This paper proposes and systematically evaluates a new normalization method for Transformer layersβ€”Peri-LN, which uses LayerNorm or RMSNorm on both the input and output of each sub-layer simultaneously.

Persistent Topological Features in Large Language Models

Yuri Gardinazzi (Area Science Park), Matteo Biagetti (Area Science Park)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Using Zigzag persistence (topological data analysis) to track the point clouds of internal representations of large language models (LLMs) across different layers, and based on this, proposing two statistical descriptors (relative birth frequency and inter-layer persistence) to characterize the dynamic evolution of prompts in the representation space.

Pfeife: Automatic Pipeline Parallelism for PyTorch

Ho Young Jhoo (Seoul National University), Nuno P. Lopes (INESC-ID Instituto Superior TΓ©cnico University of Lisbon)

CodeOptimizationComputational EfficiencyTransformerLarge Language ModelDiffusion modelImageBenchmark

🎯 What it does: Developed Pfeife, an automatic pipeline parallel tool integrated with PyTorch 2, capable of transparently capturing the complete data flow graph of models and automatically partitioning and scheduling pipeline execution across multiple GPUs.

Phase and Amplitude-aware Prompting for Enhancing Adversarial Robustness

Yibo Xu (Xidian University), Nannan Wang (Xidian University)

CodeAdversarial AttackConvolutional Neural NetworkTransformerPrompt EngineeringImage

🎯 What it does: A phase and amplitude-aware prompting method (PAP) based on frequency domain phase and amplitude spectra is proposed to enhance the robustness of models under adversarial attacks.

Pixel2Feature Attack (P2FA): Rethinking the Perturbed Space to Enhance Adversarial Transferability

Renpu Liu (Nanjing University of Information Science and Technology), Jinwei Wang (Nankai University)

CodeAdversarial AttackConvolutional Neural NetworkTransformerImage

🎯 What it does: Designed and implemented the Pixel2Feature Attack (P2FA), which transfers the perturbation space from pixel space to feature space, perturbs multiple times along the feature importance direction in feature space, and then maps the perturbation back to pixel space using feature inversion, thereby generating more transferable adversarial examples.

Plausible Token Amplification for Improving Accuracy of Differentially Private In-Context Learning Based on Implicit Bayesian Inference

Yusuke Yamasaki (NTT Corporation), Takayuki Miura (NTT Corporation)

CodeClassificationSafty and PrivacyLarge Language ModelPrompt EngineeringText

🎯 What it does: A method called Plausible Token Amplification (PTA) is proposed to generate more accurate synthetic examples in the context of differential privacy in-context learning (DP-ICL).

Playmate: Flexible Control of Portrait Animation via 3D-Implicit Space Guided Diffusion

Xingpei Ma (Guangzhou Quwan Network Technology), Shunsi Zhang (Guangzhou Quwan Network Technology)

CodeGenerationData SynthesisPose EstimationTransformerDiffusion modelVideoAudio

🎯 What it does: This paper presents Playmate, a two-stage training framework that implements audio-driven facial animation using a diffusion model guided by 3D implicit space, supporting fine-grained control of emotions and poses.

Point Cloud Dataset Distillation

Deyu Bo (National University of Singapore), Xinchao Wang (National University of Singapore)

CodeClassificationSegmentationData SynthesisKnowledge DistillationPoint Cloud

🎯 What it does: The DD3D framework is proposed, which uses a minimal amount of synthetic point clouds to replace large-scale real point clouds while maintaining the performance of classification/segmentation models.

Policy Design for Two-sided Platforms with Participation Dynamics

Haruka Kiyohara (Cornell University), Sarah Dean (Cornell University)

CodeRecommendation SystemOptimizationReinforcement LearningTabularSequential

🎯 What it does: This study investigates the engagement dynamics between audiences and content providers on two-sided platforms (such as video streaming and recruitment) and their impact on recommendation policies, proposing a recommendation strategy that considers long-term demographic effects.

Policy Gradient with Tree Expansion

Gal Dalal (NVIDIA Research), Gal Chechik (Bar-Ilan University)

CodeReinforcement LearningVideo

🎯 What it does: This paper proposes a SoftTreeMax strategy that combines tree search (Tree Expansion) with policy gradient, utilizing multi-step cumulative rewards and future state logits to form a differentiable softmax alternative.